AIAA JOURNAL

Vol. 46, No. 11, November 2008

Formulation of the k-! Turbulence Model Revisited

David C. Wilcox
DCW Industries, Inc., La Caada, California 91011

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

DOI: 10.2514/1.36541
This paper presents a reformulated version of the authors k-! model of turbulence. Revisions include the addition
of just one new closure coefcient and an adjustment to the dependence of eddy viscosity on turbulence properties.
The result is a signicantly improved model that applies to both boundary layers and free shear ows and that has
very little sensitivity to nite freestream boundary conditions on turbulence properties. The improvements to the k-!
model facilitate a signicant expansion of its range of applicability. The new model, like preceding versions, provides
accurate solutions for mildly separated ows and simple geometries such as that of a backward-facing step. The
models improvement over earlier versions lies in its accuracy for even more complicated separated ows. This paper
demonstrates the enhanced capability for supersonic ow into compression corners and a hypersonic shock-wave/
boundary-layer interaction. The excellent agreement is achieved without introducing any compressibility
modications to the turbulence model.

Ue , U1

Nomenclature
Clim
Cp
c
cf
Dk
E
e
f
H
h
k
ks , k
s

=
=
=
=
=
=
=
=
=
=
=
=

M1
p

=
=

pw , p1
Pk
PrL , PrT
ReH , Re

=
=
=
=

Re1
r
SR , SB

=
=
=

Sij

S
ij

S^ij

Tw , Taw

t
tij

=
=

stress-limiter coefcient
2
pressure coefcient,
p  p1 =
12
U1

chord length
skin-friction coefcient
dissipation of turbulence kinetic energy
Favre-averaged total energy,

e  12 ui ui  k
Favre-averaged specic internal energy
round-jet function
backward-facing-step height
Favre-averaged specic enthalpy
Favre-averaged specic turbulence kinetic energy
dimensional and dimensionless surface-roughness
height, u ks =w
freestream Mach number
mean static pressure, nite-difference-scheme
order of accuracy
surface and freestream mean static pressure
production of turbulence kinetic energy
laminar and turbulent Prandtl number
Reynolds number based on step height and
momentum thickness
Reynolds number per unit length
ne-grid point to coarse-grid point number ratio
dimensionless surface value of ! for a surface
with roughness and mass injection
Favre-averaged strain-rate tensor,
1

@ui =@xj  @uj =@xi 
2
zero-trace Favre-averaged strain-rate tensor,
Sij  13 @um =@xm ij
Galilean-invariant Favre-averaged strain-rate
tensor, Sij  12 @um =@xm ij
Favre-averaged wall temperature, adiabatic wall
temperature
time
mean viscous stress tensor

u
ui
u
vw , v
w
x
xi
xr , xs
y
, , , d
 ,  

0
0o
ij
"

, T
, w


do
ij
xy

!

ij
!
!~

= Favre-averaged boundary-layer-edge and

freestream velocity
= Favre-averaged streamwise x velocity component
= Favre-averaged velocity vector
= friction velocity
= dimensional and dimensionless vertical surface
mass-injection velocity, vw =u
= streamwise coordinate
= position vector
= reattachment and separation point
= surface-normal coordinate
= closure coefcients in the specic dissipation-rate
equation
= closure coefcients in the turbulence-kineticenergy equation
= boundary-layer thickness
= free-shear-layer spreading rate, d=dx
= value of 0 for !1 ! 0 (k-! model) or "1 ! 0
(k-" model)
= Kronecker delta
= dissipation rate
= momentum thickness
= molecular, eddy viscosity
= local and surface value of the kinematic molecular
viscosity
= mean mass density
= value of d when @k=@xi @!=@xi > 0
= Favre-averaged specic Reynolds-stress tensor
= Favre-averaged specic Reynolds shear stress
= Popes nondimensional measure of vortex
stretching parameter
= absolute value of
p
= Mean vorticity
= rotation tensor, 12
@ui =@xj  @uj =@xi 
= specic dissipation rate
= effective specic dissipation rate used to compute
eddy viscosity

I. Introduction

Presented as Paper 1408 at the 45th AIAA Aerospace Sciences Meeting

and Exhibit, Reno, NV, 811 January 2007; received 8 January 2008;
revision received 29 May 2008; accepted for publication 13 June 2008.
Copyright 2008 by David C. Wilcox. Published by the American Institute
of Aeronautics and Astronautics, Inc., with permission. Copies of this paper
may be made for personal or internal use, on condition that the copier pay the
$10.00 per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood
Drive, Danvers, MA 01923; include the code 0001-1452/08 $10.00 in
correspondence with the CCC.

President, 5354 Palm Drive. Associate Fellow AIAA.

HE k-! model was rst created independently by Kolmogorov

[1] and later by Saffman [2]. Wilcox and Alber [3] and Wilcox
[4,5] have continually rened and improved the model during the
past three decades and demonstrated its accuracy for a wide range of
turbulent ows. This paper presents the authors latest efforts aimed
at improving the model.
The new model incorporates two key modications: namely, the
addition of a cross-diffusion term and a built-in stress-limiter

2823

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

2824

WILCOX

modication that makes the eddy viscosity a function of k, !, and the

ratio of turbulence-energy production to turbulence-energy
dissipation.
The addition of cross diffusion to the ! equation was rst
suggested by Speziale et al. [6] as a remedy for the original k-!
models sensitivity to the freestream value of !. Although Speziale
et al. (as well as, for example, Menter [7], Wilcox [8], Kok [9], and
Hellsten [10]) have succeeded in using cross diffusion to eliminate
boundary-condition sensitivity, it has come at the expense of the
ability to make reasonable predictions for free shear ows. Strictly
speaking, models created in this spirit are limited in applicability to
wall-bounded ows.
Coakley [11] introduced the stress-limiter modication. Huang
[12] showed, in detail, that by limiting the magnitude of the eddy
viscosity when turbulence-energy production exceeds turbulenceenergy dissipation, this modication yields larger separation bubbles
and, most notably, greatly improves incompressible- and transonicow predictions. Kandula and Wilcox [13] veried for a transonic
airfoil that it improves predictive accuracy of the baseline k-! model
without cross diffusion and blending functions and/or nonlinear
constitutive relations such as those implemented by Menter [7] and
Hellsten [10]. In point of fact, the success achieved in this paper
demonstrates that blending functions are an unnecessary
complication.
Although these ideas are not new, the way they were implemented
is new. The k-! model was reformulated using the methodology
developed by Wilcox [14] in which boundary layers and free shear
ows are rst dissected and analyzed using perturbation methods and
similarity solutions. All aspects of the model, including boundary
conditions for rough surfaces and surfaces with mass injection, were
reformulated and validated. Then a series of computations was
performed for nearly 100 different applications, including free shear
ows, attached boundary layers, backward-facing steps, and
separated ows. The test cases cover all Mach-number ranges from
incompressible through hypersonic. Wilcox [14] presented complete
details of the models formulation, including all of the analysis,
software, input data, and experimental data used in developing and
testing the model. This paper includes results of the new k-! models
most signicant applications.

B.

Constitutive Equations

The model uses the following equations to compute the molecular

and specic Reynolds-stress tensors:
tij  2 S
ij ;

2

ij  2 T S
ij 
kij
3

1 @uk

S
ij  Sij 
3 @xk ij

k
T  ;
!~

s9
2S
ij S
ij =
!~  max !; Clim
:
 ;
8
<

Clim  78

II. New k-! Model

Mean-Flow Equations

The Favre-averaged equations for conservation of mass,

momentum, and energy are as follows:
@
@




ui   0
@x
@t
i

(1)

@
@p
@
@


ui  


uj ui   


tji 
ji 
@x
@x
@x
@t
j
i
j

(2)






1
@
1
@

e  ui ui  k 

uj h  ui ui  k
2
@xj
2
@t



 


@

T @h
@k

k

u
t 
ij  

 
@xj i ij
PrL PrT @xj
! @xj
(3)
Note that the energy-conservation equation (3) ensures conservation
of total energy E, which includes the kinetic energy of the turbulence.

(4)

(5)

(6)

(7)

The stress-limiter modication [Eq. (6)] uses the zero-trace version

of the mean strain-rate tensor (viz., S
ij ). Some turbulence-model
researchers prefer the magnitude of the vorticity vector in place of

2S
ij S
ij 1=2 . Using the magnitude of the vorticity with Clim  0:95 is
satisfactory for shock-separated-ow predictions up to Mach 3 (and
possibly a bit higher). However, numerical experimentation with this
k-! model has shown that it has a detrimental effect on hypersonic
shock-induced separation, some (but not all) attached boundary
layers, and some free shear ows (especially the mixing layer).
C.

Turbulence Model Equations

The equations governing the turbulence kinetic energy and

specic dissipation rate are
@
@


k 


uj k
@t
@xj

For the sake of clarity, this paper will refer to the reformulated k-!
equations as the new k-! model. This paper focuses on whats new
about the model relative to previous versions. Readers interested in
all aspects of the model and its development can nd a complete
presentation by Wilcox [14].
A.

Consequently, the equations diffusion term includes the explicit

appearance of the molecular and turbulent diffusion of k.


ij

@ui
@
 
k! 
@xj
@xj



 

 

k @k
! @xj

@
@
!
@u


! 


uj !  
ij i  
!2
@t
@xj
k
@xj


 

@k @!
@

k @!

 d

! @xj @xj @xj
! @xj

(8)

(9)

The turbulence-kinetic-energy equation (8) contains no special

compressibility terms involving pressure work, diffusion, or
dilatation. Although a dilatationdissipation modication to the k
equation improves compressible mixing-layer predictions (see
Wilcox [14]), the same modication has a detrimental effect on
shock-separated-ow predictions. Hence, it is omitted from the k
equation for general applications.
Note that the turbulent-diffusion terms in Eqs. (8) and (9) (i.e., the
terms proportional to   and ) are proportional to
k=! rather than
to the eddy viscosity. This means that the only terms in these
equations that are implicitly affected by the stress limiter are the
production terms (via the Reynolds-stress tensor). Consequently, the
new k-! model can serve as the foundation of a model with a more
general prescription for computing the Reynolds-stress tensor such
as an algebraic stress model, a full stress-transport model, and even a
detached eddy simulation. In principle, there should be no need to
revise the models closure coefcients to accommodate an
alternative way of computing the Reynolds stresses. Wilcox [14]
demonstrated this exibility for a stress-transport model.

2825

WILCOX

D.

Closure Coefcients

The various closure coefcients appearing in the new k-! model

are
;
  13
25

9
  100
;

(
d 

  o f  ;

PrT  89 (10)

SB 
0;
do ;

@k
@xj
@k
@xj

@!
@xj
@!
@xj

0
>0

o  0:0708;



ij jk S^ki 
;


!  

 !3 
Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

   35;

  12;

1
do 
8

f 

(11)

1  85
!
1  100
!

1 @um
S^ki  Ski 

2 @xm ki

(12)

(13)

The round-jet parameter

! is computed with S^ki , which, unlike the
compressible strain rate recommended by Papp and Dash [15], is
Galilean-invariant. This is necessary because using Ski or S
ki yields
undesired effects in two-dimensional compressible ows.
E.

Boundary Condition for Rough and Smooth Surfaces

For surfaces that include surface roughness, the model uses the noslip condition for velocity and k. The surface value of ! depends
upon the dimensionless surface-roughness height k
s . The boundary
condition for ! is
!

u2
S
 R

at y  0

(14)

where SR was chosen to provide a close match to measured roughsurface boundary-layer data of Nikuradse, as noted in [16]. The
following correlation between SR and k
s reproduces measured
effects of sand-grain roughness for values of k
s up to about 400:
( 200 2

k  ;
k
s  5
SR  100s
(15)

200 2
100 5ks

k   k
e
; k
s >5
k
s

A surface is considered to be hydraulically smooth when k

s < 5.
For such surfaces, we can combine Eqs. (14) and (15) to obtain the
slightly-rough-surface boundary condition for !: namely,
!

40; 000
k2s

at y  0

(16)

Because the turbulence-model solution for a hydraulically smooth

surface is nearly identical to the perfectly smooth-surface solution,
Eq. (16) can be used for smooth surfaces, with ks chosen to insure that
k
s < 5.
F.

where the value of SB was chosen to achieve optimum agreement

between measured [17] and computed velocities. The correlation
between SB and dimensionless mass-injection velocity v
w is given in
analytical form as

Boundary Condition for a Surface with Mass Transfer

For a surface with mass transfer, we again implement the no-slip

condition for the mean velocity and k. When the surface has blowing
corresponding to vw > 0, the boundary condition for ! is
!

u2
S
 B

at y  0

Table 1
Flow
Far wake
Mixing layer
Plane jet
Round jet
Radial jet

(17)

25
v

1
 5v
w
w

(18)

When the surface has suction corresponding to vw < 0, the value of !

appropriate for a smooth surface [Eq. (16)] should be used.

III. Cross Diffusion

One of the key modications in the new k-! model is addition of a
cross-diffusion term. The term proportional to d in Eq. (9) is referred
to as cross diffusion. It depends upon gradients of both k and !.
A.

Free Shear Flows

In free shear ows the cross-diffusion term enhances production of

!, which in turn increases dissipation of k (assuming d > 0). This
occurs for small freestream values of k and !, for which both
quantities decrease approaching the shear-layer edge. The overall
effect is to reduce the net production of k, which reduces the
predicted spreading rates.
Several authors, including Speziale et al., [6] Menter [7], Wilcox
[8], Peng et al. [18], Kok [9], and Hellsten [10], have attempted to
improve the k-! model by adding cross diffusion. Although all have
achieved some degree of success in wall-bounded ows, the models
are far less realistic for free shear ows. Inspection of Table 1 shows
that spreading rates predicted by such models differ signicantly
from measured values.
Menter [7] and Hellsten [10] enjoyed more success with cross
diffusion than Speziale et al. [6] and Peng et al. [18]. Both introduced
blending functions that cause all of the models closure coefcients
to assume values appropriate for the k-! model near solid boundaries
and to asymptotically approach values similar to those used with the
k-" model [19] otherwise. The net result is a model that behaves very
much like the Wilcox [4] k-! model for wall-bounded ows and
more like the k-" model for free shear ows.
Wilcox [8] and, more recently, Kok [9] tried a similar concept with
the cross-diffusion coefcient d , given by
(
@k @!
0;
0
@xj @xj
d 
(19)
@k @!
do ; @xj @xj > 0
Additionally,   assumes a value larger than 12. It is important to
suppress the cross-diffusion term close to solid boundaries for wallbounded ows. This is true because, as discussed in detail by Wilcox
[14], cross diffusion changes the near-surface structure of the !
equation in a way that undermines sublayer predictions. Just as
Menters blending function causes d to approach zero near a solid
boundary, so does Eq. (19), because k increases and ! decreases in
the viscous sublayer. Although simpler than Menters blendingfunction approach, Wilcox [8] and Kok [9] chose values for do that
yield free-shear-layer spreading rates that are farther from
measurements than those predicted by the k-" model. Specically,
Wilcox [8] set do  103 ,   35, and    1, whereas Kok [9] opted
for do    12 and    23.
However, other values of the k-! models closure coefcients
exist that yield closer agreement with measured spreading rates. Note

Two-equation model free-shear-ow spreading rates

Speziale [6]

Peng [18]

Kok [9]

Wilcox [4]

New k-!

Measured

0.221
0.082
0.089
0.102
0.073

0.206
0.071

0.096
0.040

0.191
0.056
0.083
0.107
0.068

0.496
0.141
0.135
0.369
0.317

0.326
0.096
0.108
0.094
0.099

0.3200.400
0.1030.120
0.1000.110
0.0860.096
0.0960.110

2826

WILCOX

rst that, based on the analysis of a turbulent front by Lele [20], there
are two necessary conditions for the front to propagate. Specically,
we must have

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

do >    

and

  > do

(20)

Figure 1 shows how predicted spreading rates vary with do for the
far wake, the mixing layer, and the plane jet. The curves shown were
computed with all other closure coefcients as specied in Eqs. (10)
and (12). To isolate effects of cross diffusion, results shown
correspond to having no stress limiter: that is, !~  ! in Eq. (6). The
limiter has virtually no effect on the far wake and the plane jet. It
reduces the mixing-layer spreading rate by less than 6%. Of greatest
relevance to the present discussion, the value of   is 3=5. As shown,
spreading rates for all three cases are greatest when do is equal to its
minimum permissible value according to Eq. (20) (viz.,
do     ). Predicted values decrease monotonically as do
increases and fall below the lower bound of measured spreading rates
for all three cases when do  15, which is much less than the
maximum allowable value of 35.
Figure 2 shows how predicted spreading rates vary with   when
we set do equal to its minimum permissible value. As noted,
computations were done with all closure coefcients other than   ,
as specied in Eqs. (10) and (12) in the absence of the stress limiter.
Computed spreading rates for all three cases decrease monotonically
as   increases. Computed 0 values lie above the range of measured
0 for all three cases when   < 0:55 and below when   < 0:70.
Thus, we conclude that
0:55 <   < 0:70

for   12

(21)

These results provide the rationale for selecting    35 and do  18

for the new k-! model.

Fig. 1

Fig. 2

B.

Round-Jet/Plane-Jet Anomaly

Inspection of Table 1 shows that, with the exception of the new

k-! model, all of the turbulence models listed predict that the round
jet spreads more rapidly than the plane jet. Measurements indicate
the opposite trend, with the round-jet spreading rate being about 10%
lower than that of the plane jet. This shortcoming, common to most
turbulence models, is known as the round-jet/plane-jet anomaly.
Pope [21] proposed a modication to the " equation that resolves
the round-jet/plane-jet anomaly for the k-" model [19]. In Popes
modication, the dissipation of dissipation term in the " equation is
replaced by
C"2

"2
"2
! C"2  C"3
p

k
k

(22)

where C"2 and C"3 are closure coefcients. In terms of k-! model
parameters, " / k!. The parameter
p is a nondimensional measure
of vortex stretching dened as

p 

ij jk Ski

"=k3

(23)

Popes [21] reasoning is that the primary mechanism for transfer of

energy from large to small eddies is vortex stretching. Any
mechanism that enhances vortex stretching will increase this rate of
transfer. Because the energy is being transferred to the smallest
eddies in which dissipation occurs, the dissipation must necessarily
increase. Because mean-ow vortex lines cannot be stretched in a
two-dimensional ow,
p is zero for the plane jet. By contrast, the
vortex-stretching parameter is nonzero for an axisymmetric mean
ow. As argued by Pope, this corresponds to the fact that vortex rings
are stretched radially. Thus, we expect to have
p 0 for a round jet.
Using C"3  0:79 reduces the k-" models predicted spreading
rate to 0.86, consistent with measurements. However, as pointed out
by Rubel [22], the Pope [21] correction has an adverse effect on

Effect of cross diffusion on free-shear-ow spreading rates for

 35 and
 12; shaded areas depict measured-value ranges.

Effect of cross diffusion on free-shear-ow spreading rates for
do 


and
 12; shaded areas depict measured-value ranges.

2827

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

WILCOX

model predictions for the radial jet, which also has nonzero
p .
Without the Pope correction, the k-" model predicts a radial-jet
spreading rate of 0.094, which is close to the measured range of 0.096
to 0.110 (see Tanaka and Tanaka [23] and Witze and Dwyer [24]).
Using the Pope [21] correction for the radial jet reduces the k-"
model-predicted spreading rate to 0.040. Hence, as noted by Rubel
[22], the round jet/plane jet anomaly has been exchanged for a round
jet/radial jet anomaly.
In contrast to the k-" model, as indicated in Table 1, the Wilcox [4]
k-! model predicts comparable spreading rates for both the round
and radial jets, both larger than the predicted plane-jet spreading rate.
Similarly, when a constant value of   0:0708 is used for the new
k-! model, the predicted round- and radial-jet spreading rates are
0.177 and 0.168, respectively. Numerical experimentation shows
that if  is reduced to 0.06, the models spreading rates for both the
round and radial jets are close to the measured values. Because
Popes [21] argument implies nothing regarding the functional
dependence of the modication upon
p , it is completely consistent
to propose that  depends upon this parameter in a manner that
reduces the value of  as needed for both ows. Thus, as a
generalization of the Pope modication, the reformulated k-! model
uses the following prescription for :
  o f 

(24)

where
o  0:0708;

f 

1  85
!
1  100
!

(25)

and


ij jk Ski 


!  

 !3 

(26)

Comparison of Eqs. (23) and (26) shows that

!  j
p j. Also, the
functional form of f is such that its asymptotic value is 0.85, so that
  0:06 for large values of
! . Finally, note that the vortexstretching parameter normally is very small in axisymmetric
boundary layers because ! is very large.
Although the usefulness of Popes [21] correction as represented
by Eqs. (22) and (23) is limited by a aw in the k-" model, the
concepts underlying the formulation are not. We can reasonably
conclude that Popes analysis provides a sensible reection of the

physics of turbulent jets, at least in the context of !-based twoequation models.

C.

Computed and Measured Velocity Proles

Figure 3 illustrates the remarkable effect that a modest amount of

cross diffusion has upon free-shear-ow results. For reference, the
gure includes results obtained for the standard k-" model [19] and
for the renormalization group (RNG) k-" model [25]. Experimental
data for the far wake are from Fage and Falkner [26] and Weygandt
and Mehta [27], whereas those for the radially spreading jet are from
Witze and Dwyer [24].
D.

Sensitivity to Finite Freestream Boundary Conditions

Two-equation models have a unique and unexpected feature when

nonzero freestream boundary conditions are specied for k, !, ", etc.
Specically, even if we select k and the second turbulence property
(!, ", etc.) to be sufciently small that both k and T are negligible,
the solution is sensitive to our choice of the second turbulence
propertys freestream value. This is an important consideration
because most computations are done with these assumptions.
Figure 4 shows how the spreading rate 0 =0o varies with the
freestream value of ! for the new k-! model and the Wilcox [4] k-!
model for the far wake, the mixing layer, and the plane jet. It also
shows the variation of 0 with the freestream value of " for the
standard k-" model [19]. All computations were done with the same
(very small) dimensionless eddy viscosity.
All three models predict a decrease in spreading rate as the
freestream value of ! or " increases. In all three graphs, the
freestream value is scaled with respect to the value at y  0, which is
either equal to or very close to the maximum value for each ow. As
shown, without cross diffusion, the k-! model displays a strong
sensitivity to the freestream value of !. The addition of cross
diffusion greatly reduces the sensitivity. The k-" model predicts very
little sensitivity to the freestream value of ". The graphs also show
that if the freestream value is less than 1% of the maximum value
[!1 =!
x; 0 < 0:01, "1 ="
x; 0 < 0:01], there is virtually no effect
on the predicted spreading rate. This is certainly not an unreasonable
constraint, because using a freestream value of ! or " in excess of 1%
of the peak value would very likely correspond to using a physically
unrealistic value.
There is no mystery about why the solution should have some
sensitivity to freestream boundary conditions. We are, after all,
solving a two-point boundary-value problem, which requires

Fig. 3 Far-wake and radially spreading-jet solutions: new k-! model (solid line), k-" model [19] (dashed line), RNG k-" model [25] (dotted line), and
measured [24,26,27] (open and lled circles).

2828

WILCOX

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

Fig. 4 Sensitivity of free-shear-ow spreading rates to freestream conditions: new k-! model (solid line), Wilcox [4] k-! model (dashed line), and k-"
model [19] (dotted line).

freestream boundary conditions on all variables, including ! and ".

In light of this, it is clear that there must be some range of boundary
values that affect the solution. Figure 4 shows that there is a welldened limiting form of the solution for vanishing freestream
boundary values.
It is the odd nature of the differential equation for " that makes the
k-" model much less sensitive to freestream conditions than the k-!
model. Specically, because its dissipation term is proportional to
"2 =k, the equation is singular as k ! 0 for nite freestream values of
". This unusual behavior of the " equation obviates the need to invest
enough thought to avoid prescribing physically unrealistic
freestream values for a quantity such as ". Although this may be
comforting to engineers who do not care to invest such thought,
turbulent-uid-ow applications exist [14] for which being sloppy
with freestream boundary conditions can foil the protection provided
by the " equation.
Studies have been published [28] in which the freestream value of
! has been set to very large values. With an extremely large
freestream !, any k-! model solution for many ows, especially free
shear ows, will be grossly distorted. This type of analysis is very
misleading because having freestream values of ! more than a
percent or so of the maximum value in the turbulent region is
physically incorrect. What ! quanties is the vorticity of the energycontaining eddies. Assigning huge values of ! in the freestream
would imply that there is signicant uctuating vorticity above the
turbulent region, which is absurd.
As an analogy, consider the laminar boundary layer with zero
pressure gradient. The boundary-layer equations admit a similarity
solution (viz., the Blasius solution). Imagine that rather than
imposing the freestream boundary condition on the velocity, we
specify the freestream value of the vorticity. For zero freestream
vorticity, the solution is identical to the Blasius solution. Figure 5
shows how the skin friction varies with the freestream vorticity 1 .
There is signicant distortion when 1 exceeds a 100th of a percent

(0.01%) of the peak vorticity o in the boundary layer. How different

is this from selecting a physically unrealistic freestream boundary
condition on the vorticity of the energy-containing eddies with the
k-! model? The same logic that would cite the sensitivity to a
freestream value of ! that exceeds 1% of the peak value in the
turbulent region as a aw in the turbulence model would conclude
that Prandtls boundary-layer equations are fundamentally awed
for the same reason!

E.

Attached Boundary Layers

As demonstrated by Kok [9], cross diffusion does not necessarily

cause a loss of accuracy in predicting effects of pressure gradient on
attached boundary layers. This is true of the new k-! model. Figure 6
compares computed and measured skin-friction and velocity proles
for two attached boundary layers with a strong adverse pressure
gradient. The graphs to the left correspond to the classic Samuel
Joubert experiment [29], which has an increasingly adverse gradient.
This is an important test case that was poorly predicted by virtually
all turbulence models at the 198081 AFOSR-HTTM-Stanford
Conference on Complex Turbulent Flows [29]. The graphs to the
right correspond to the incipient-separation case of Stratford [30]. To
this authors knowledge, this prediction is the closest to
measurements of any turbulence model used to predict the ow.
Virtually all k-" models, for example, predict skin friction that is 4
times the measured value.
The new k-! model, as with previous versions, applies equally
well to supersonic and even hypersonic boundary layers. Figure 7
compares computed and measured velocity proles for Mach 4.5 and
Mach 10.3 at-plate boundary layers [31]. Note that U  u =u ,
where u is the van Driest (see Wilcox [14]) scaled velocity.

F.

Turbulent/Nonturbulent Interfaces

More often than not, turbulence-model equations that are in

general usage appear to predict sharp interfaces between turbulent
and nonturbulent regions (i.e., interfaces in which discontinuities in
derivatives of ow properties occur at the edge of the shear layer). As
noted by Wilcox [14], these interfaces bear no relation to the physical
turbulent/nonturbulent interfaces that actually uctuate in time and
have smooth Reynolds-averaged properties. Omitting details of the
analysis for the sake of brevity, for the k-! model with cross diffusion
included (but no stress limiter), the asymptotic behavior of u, k, and !
approaching a turbulent/nonturbulent interface is given by
9
Ue  u  uo
1  y=nu =
k  ko
1  y=nk
;
!  !o
1  y=n!
Fig. 5 Effect of freestream vorticity on an incompressible laminar atplate boundary layer.

as

y!

(27)

where uo , ko , and !o are integration constants and the three

exponents are

2829

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

WILCOX

Fig. 6 Attached boundary layers with strong adverse pressure gradient: computed values (solid line) and Samuel and Joubert experiment [29] (circles)
and Stratford [30] (circles).

Fig. 7 Supersonic and hypersonic boundary layers: Coles experiment [31] (circles) and Watson experiment [31] (circles).

9

>
nu   do >
=
nk  do
  do >
>
;
n!  
 
do

(28)

With a stress limiter included, nk and n! are unchanged, but the

solution for the velocity is such that nu  nk .
For the solution to give u ! Ue , k ! 0, and ! ! 0 as we
approach the turbulent/nonturbulent interface from the turbulent
side, all three exponents in Eqs. (28) must be positive. This is true
provided the closure coefcients ,   , and do satisfy the following
constraints:
do >    

and

  > do

(29)

These are identical to the constraints deduced by Lele [20] in

analyzing a turbulent front. Table 2 lists the values of the exponents
for several k-! models, each having unique behavior.
1) Hellstens [10] model features continuous second derivatives
for u, k, and !, so that its weak-solution behavior should be of no
consequence in a second-order-accurate numerical solution.

Table 2 Turbulent/nonturbulent interface exponents for k-! models

Model
Hellsten [10]
Kok [9]
Menter [7]
New k-!



do

nu

nk

n!

1.000
0.500
0.856
0.500

1.100
0.667
1.000
0.600

0.400
0.500
1.712
0.125

3.333
1.000
0.546
20

3.333
1.500
0.546
20

2.333
0.500
0:454
19

2830

WILCOX

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

Fig. 8 Computed and measured velocity defect near the boundarylayer edge for a at-plate boundary layer using three k-! models: new
(solid line), Kok (dashed line), Hellsten (dotted line), Klebanoff (circles),
Wieghardt and Tillman [33] (squares), and Winter and Gaudet [34]
(triangles).

2) Koks [9] model has classic weak-solution behavior with

discontinuities in the slope of u and !.
3) Because Menters [7] model fails to satisfy the second condition
of Eq. (29), the solution for ! approaches 1 as y ! .
4) The new k-! model is analytic approaching the interface so that
it does not have a nonphysical weak-solution behavior.
Hellsten [10] made the case for choosing the values of the models
closure coefcients based on achieving smooth solution behavior at a
turbulent/nonturbulent interface. Part of Hellstens arguments
include a claim that to achieve such behavior it is necessary to have
  > 1. Because the new k-! model has a completely analytical
solution at such an interface while having   < 1, a closer look is in
order.
Figure 8 compares computed and measured [3234] velocity
proles in the immediate vicinity of the boundary-layer edge for a
constant-pressure boundary layer. Hellsten [10] presented a similar
graph showing the linear approach of Koks [9] velocity prole and
the discontinuity in slope at the interface. By contrast, both the
Hellsten [10] model and the new k-! model exhibit a smooth
approach to freestream values, with both curves falling within
experimental-data scatter.
The apparent contradiction in Hellstens [10] claim regarding the
minimum value of   needed to achieve smooth solutions near a
turbulent/nonturbulent interface is easily resolved. Inspection of
Fig. 8 shows that below y=  0:95 all three velocity proles are very
nearly linear functions of y=. The region in which the asymptotic
solution given in Eqs. (27) and (28) is valid lies well within the upper
15% of the boundary layer, depending on the precise values of nu ,
nk , and n! . Consequently, on the scale shown in the graph, it is
difcult to discern much difference between the solutions for the
Hellsten model and the new k-! model. As noted, both models have
continuous second derivatives (and higher) approaching the
interface and should be expected to cause no troublesome numerical
issues to arise.

IV.

Stress Limiter

The second key modication in the new k-! model occurs in the
expression for the eddy viscosity. In the new model, T is the ratio of
k to ! multiplied by a factor that is, effectively, a function of the
turbulence-kinetic-energy production-to-dissipation ratio. This
modication greatly improves the models predictions for supersonic
and hypersonic separated ows.
Note that this modication pertains to the proposed constitutive
relation between the Reynolds stresses and mean-ow properties,
rather than to the k-! model per se. The virtues of the stress limiter
(often referred to as a weakly nonlinear stress/strain-rate relationship) can be realized by using a nonlinear stress/strain-rate
relationship or even by computing the Reynolds stresses with a
stress-transport model based on the k and ! equations. As noted

earlier, because the stress limiter appears in the k and ! equations

only through the Reynolds-stress tensor, the new k-! model can be
used, without modication, with other prescriptions for the Reynolds
stresses. Wilcox [14], for example, presented complete details for a
k-!-based stress-transport model.
Coakley [11] was the rst to suggest that shock-separated ows
can be more accurately simulated with the k-! model by simply
limiting the magnitude of the Reynolds shear stress when production
of turbulence kinetic energy exceeds its dissipation. He developed a
stress limiter that showed some promise for improving k-! model
predictions. Menter [7], Kandula and Wilcox [13], Durbin [35], and
Huang [12], for example, subsequently conrmed the effectiveness
of a stress limiter for ow speeds up to the transonic range.
Durbin [35] and Moore and Moore [36] assessed the realizability
of turbulence-energy production predicted using the Boussinesq
approximation. They observed that for ows such as impinging jets
and the inviscid highly strained ow approaching a stagnation point,
without the assistance of a stress limiter, the Boussinesq
approximation leads to unrealistically high turbulence-energy
levels: levels that are not realized in nature. Moore and Moore
proposed the following general relation for limiting the Reynolds
~ they proposed that !~ is given by
stress. Letting T 
k=!,
8
s9
<
21 S
ij S
ij  22 ij ij =
!~  max !; C0 !  Clim
(30)
:
;

1  2 
Table 3 lists the values of the constants C0 , Clim , 1 , and 2 proposed
by several researchers.
To understand the way in which the stress limiter suppresses the
magnitude of the Reynolds shear stress, we rst simplify Eq. (30) for
the most commonly used version that has C0  0, 1  1, and
2  0: namely,
8
s9
<
2S
ij S
ij =
!~  max !; Clim
(31)
:
 ;
In a shear layer, we know that 2S
ij S
ij 
@u=@y2 . So Eq. (31) tells
us that


p
@u

k @u

 min
; Cl1


k
(32)

xy  T
lim
@y
! @y
Also, observe that in the absence of a stress limiter, the ratio of
production to dissipation in the equation for turbulence kinetic
energy is


Pk


k=!
@u=@y2
@u=@y 2



p
(33)

Dk

k!
 !
Thus, the stress-limiter modication is such that

xy  Cl1
lim

p

k for

Pk
Cl2
lim
Dk

(34)

Consequently, the stress limiter drives the Reynolds shear stress

toward the form Bradshaw et al. [37] implemented in their oneequation turbulence model. When Clim  1, the coefcient
1

 1=2  0:30, which matches the value of Bradshaws
Clim
Table 3

Stress-limiter coefcients

Reference
Coakley [11]
Durbin [35]
Menter [7]
Moore and Moore [36]
New k-!

C0

Clim

0
0
0
2.85
0

1.00
1.03
1.00
0.75
0.88

1
1
0
1
1

0
0
1
1
0

2831

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

WILCOX

constant [14]. For the new k-! model, we nd that

1

 1=2  0:34.
Clim
Interestingly, in a shear layer the turbulence-kinetic-energy
production term in the Saffman and Wilcox [38] k-!2 model is
Pk  0:30
kj@u=@yj. Hence, production of k is constrained although
the eddy viscosity is not. This is the reason that Wilcox and Traci [39]
were able to accurately compute the increase in turbulence kinetic
energy approaching a stagnation point. This is not possible with a
two-equation turbulence model that does not implement a stress
limiter (Durbin [35]), because the strain-rate eld is such that Pk =Dk
is typically in excess of 100. Although experimental data are not
shown in Fig. 9, the computed amplication is consistent with the
measurements of Bearman [40].
Figure 10 shows the dramatic improvement in predicted pressure
coefcient for Mach 0.8 ow past a NACA 0012 airfoil at a 2.26 deg
angle of attack [13]. The solid curves identied as original
correspond to the Wilcox [4] k-! model, which does not use a stress
limiter. The dashed curves identied as SST correspond to the same
model with a stress limiter applied using Clim  1. The most dramatic
difference is the location of the shock. Without the stress limiter, the
predicted shock location is farther downstream than the measured
location. Adding the stress limiter increases the size of the separation
bubble on the upper surface of the airfoil, causing the computed
shock location to lie much closer to the experimentally observed
location.
The following subsections compare computed and measured ow
properties for separated ows with ow speeds from incompressible
to hypersonic. All computations were done using a second-orderaccurate NavierStokes solver developed by MacCormack [41]. In

every case, generalized Richardson extrapolation was performed and

Appendix A summarizes the results.
A.

Incompressible Backward-Facing Steps

We rst consider the backward-facing step of Driver and

Seegmiller [42]. Figure 11 compares computed and measured skinfriction and surface-pressure coefcients for the new k-! model. The
gure also includes values predicted by the Wilcox [4] k-! model to
help discern the effect of the stress limiter. With the exception of the
reattachment point, all computed ow properties are nearly identical.
The only signicant difference is the reattachment length, which is
13% longer with the stress limiter. Menter [7] found a similar effect in
his computations.
Flow past a backward-facing step is mildly dependent on
Reynolds number. As summarized by Jovic and Driver [43],
reattachment length is somewhat shorter at low Reynolds numbers.
To assess the effect of Reynolds number on k-! model backwardfacing-step predictions, we now consider the case documented by
Jovic and Driver [44]. Reynolds number based on step height for the
JovicDriver backward-facing-step experiment is ReH  5000. By
contrast, the considered DriverSeegmiller case has ReH  37; 500.
Figure 12 compares computed and measured skin-friction and
surface-pressure coefcients. Both versions of the k-! model predict
cf and Cp variations that fall within a few percent of measured values
over most of the oweld. Predicted reattachment length is 6:64H (a
7% increase over the ReH  37; 500 prediction) for the Wilcox [4]
k-! model and 7:28H (a 3% increase) for the new k-! model.
Because the measured length is 6:00H (a 4% decrease), neither
model reects the measured reduction of recirculation-region length.
These two examples show that using the stress limiter with the k-!
model increases the size of the separated region. The stress limiter
increases differences between predicted and measured reattachment
length for ow past backward-facing steps (Figs. 11 and 12). This is
understandable because the model yields reattachment lengths that
are very close to measured lengths in the absence of the stress limiter.
To gain some insight into the stress limiters nature, recall that we
compute the eddy viscosity according to Eq. (6). In implementing the
stress-limiter concept for his hybrid k-!=k-" model, Menter [7]
selected Clim  1 and excluded it from the hybrid !=" equation.
Durbin [35] recommended Clim  1:03 for use with a pure k-!
model.
Figure 13 indicates how reattachment length for the ReH 
37; 500 backward-facing step varies with Clim . As shown, xr
increases in a monotone fashion as Clim increases. The asymptotic
value in the absence of a stress limiter (i.e., for Clim  0) is

Fig. 9 Variation of turbulence kinetic energy approaching a stagnation

point; Saffman-Wilcox [38] k-!2 model (solid lines).

Fig. 10 Comparison of computed and measured surface pressure for

transonic ow past a NACA 0012 airfoil at a 2.26 deg angle of attack.

Fig. 11 Computed and measured skin friction and surface pressure for
ow past a backward-facing step; ReH  37; 500: new k-! model (solid
line) Wilcox [4] k-! model (dashed line), and Driver-Seegmiller [42]
(circles).

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

2832

WILCOX

Fig. 12 Computed and measured skin friction and surface pressure for
ow past a backward-facing step; ReH  5000; new k-! model (solid
line), Wilcox [4] k-! model (dashed line), and Jovic and Driver [44]
(circles).

Fig. 13 Effect of the stress-limiter coefcient on computed reattachment length for a backward-facing step with ReH  37; 500.

xr  6:33H, which is 1% larger than the measured value. Selecting

Clim  7=8 yields a value of xr  7:07H, which is within 13% of the
measured length.
B.

Fig. 14 Application of several turbulence models to transonic ow past

an axisymmetric bump: new k-! (solid line), Wilcox [4] k-! (dashed
line), Menter [7] k-!=k-" (dotted line), SpalartAllmaras [46] (wide
dashed line), and Bachalo and Johnson [45] (circles).

long-dashed curve corresponds to the SpalartAllmaras [46]

one-equation model. Although separation-shock location and
separation are about the same as for the Wilcox [4] model, computed
Cp is closer to measured Cp near reattachment. The dotted curve
corresponds to Menters [7] k-!=k-" model with a stress limiter.
Computed and measured shock locations and Cp are quite close.
Using Clim  1 with the new k-! model yields Cp nearly identical
to the Menter [7] prediction. But the improved results for this ow
come at the expense of much greater discrepancies between theory
and experiment for both smaller and larger Mach numbers. This
explains why Menters model, which appears to be ne-tuned for the
transonic regime, fares well for Mach numbers from incompressible
up to transonic speeds, but very poorly for supersonic and hypersonic
ows. The primary culprit is not so much the stress-limiter strength,
as reected by the value of Clim , as it is the Boussinesq
approximation. By accepting 7% discrepancies between predicted
and measured properties for this ow, which are comparable to those
obtained with the SpalartAllmaras model, the new k-! model
reproduces measurements quite closely, all the way from
incompressible speeds to the hypersonic regime.

Transonic Flow Over an Axisymmetric Bump

The transonic-bump experiment of Bachalo and Johnson [45] is a

particularly challenging separated-ow application. In the experiment, a long slender bump is fared onto the surface of a cylinder. The
freestream Mach number is M1  0:875, the unit Reynolds number
is Re1  4  106 ft1 , and the surface is adiabatic. A shock wave
develops over the bump, which separates the boundary layer. The
ow reattaches in the wake of the bump, giving rise to a reattachment
shock. This ow is very difcult to predict because the bump surface
pressure is extremely sensitive to the size of the separation bubble,
which is strongly coupled to the precise shock locations.
Figure 14 compares computed and measured Cp for four
turbulence models. The short-dashed curve corresponds to the
Wilcox [4] k-! model, which does not have a stress limiter. Although
the predicted separation-shock location differs from the measured
location by only 6% of the bumps chord length, computed and
measured Cp differ signicantly. The solid curve corresponds to the
new k-! model, which includes a stress limiter. Differences between
computed and measured Cp are generally less than 7%. The

Fig. 15 Effect of the stress limiter on the k-! model for Mach 2 ow past
a backward-facing step: with limiter (solid line), without limiter (dashed
line), and Samimy et al. [47] (circles).

WILCOX

2833

Fig. 16 Schematics of supersonic ow into a compression corner and shock-wave/boundary-layer interaction (reecting shock).

C.

Mach 2 Flow Past a Backward-Facing Step

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

We turn now to compressible ow past a backward-facing step.

The case we will discuss has a freestream Mach number of 2.07, the
incident boundary layer has a momentum-thickness Reynolds
number of Re  1:2  104 , and the surface is adiabatic. This ow
was investigated experimentally by Samimy et al. [47]. The

computation was done with the new k-! model with and without the
stress limiter.
As shown in Fig. 15, with Clim  7=8, the stress limiter has a
barely noticeable effect on the computed surface-pressure
coefcient. Computed and measured values of Cp differ by less
than 7% for the entire oweld. Predicted reattachment length with

Fig. 17 Comparison of computed and measured surface pressure and skin friction for Mach 3 shock-separated ows using the new k-! model: with
limiter (solid line), without limiter (dashed line), Settles et al. [50] (circles), Dolling and Murphy [51] (squares), Reda and Murphy [52] (diamonds), and
Murthy and Rose [53] (circles).

2834

WILCOX

the limiter is xr  2:67H. The length decreases to xr  2:55H

without the limiter. Both values are within a few percent of the value
measured by Samimy et al. [47], which is xr  2:76H. Using Clim 
1 for this ow increases xr to 2:78H, which is also quite close to the
measured reattachment length. Clearly, the effect is less pronounced
than for an incompressible backstep. However, as we will see in the
next subsection, with Clim  1 the stress-limiter effect is far too
strong at Mach 3.

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

D.

Mach 3 Compression Corners and Reecting Shocks

Supersonic ow into a compression corner and reection of an

oblique shock from a at surface have proven to be the most
challenging of all two-dimensional separated-ow applications.
Figure 16 sketches these two geometries, including some of the main
features of the ow structure for each. Although the geometries are
fundamentally different, these ows are nevertheless very similar.
Through extensive experimental investigations, Petrov et al. [48] and
Chapman et al. [49] developed the free-interaction concept. They
found that ow details in the vicinity of separation are local and
depend almost entirely on Mach number and static-pressure ratio
across the separation shock. Thus, if we test a model for
compression-corner ows, we should simultaneously test the model
for reecting shocks to check consistency with the free-interaction
concept.
Figure 17 compares the computed and measured surface pressure
and skin friction for two compression-corner ows and a reectingshock case. All three ows have a freestream Mach number close to 3
and have separation bubbles of different sizes. The two compressioncorner ows have wedge angles of 20 and 24 deg, corresponding to
experiments conducted by Settles et al. [50] and by Dolling and
Murphy [51]. Both cases have a wall to adiabatic-wall temperature
ratio Tw =Taw  0:88, corresponding to very mild cooling. The
reecting-shock case was investigated experimentally by Reda and
Murphy [52] and by Murthy and Rose [53]. The incident shock
makes an angle of 31 deg with the horizontal and turns the ow by
13 deg. The surface for this case is adiabatic.
The graphs include results for the new k-! model with and without
the stress limiter. In all three cases, with the stress limiter, computed
and measured surface pressures are very close. Most important, the
initial pressure rise in the computed owelds matches the measured
rise. This means that the separation shock is in the same location in
the numerical and experimental owelds. The predicted pressure
plateau in the separation bubble and skin friction downstream of
reattachment is close to measurements. Discrepancies between
computed and measured cf downstream of reattachment indicates
that the rate of recovery from separation and the return to equilibrium
conditions is a bit different.
Without the stress limiter, the computed separation-shock location
is clearly further downstream than measured, which distorts the
entire oweld.
The similarity between the shapes of the computed pw =p1 and cf
curves for the shock-wave/boundary-layer interaction and the 24 deg
compression-corner ow is striking. Because the overall pressure

rise is nearly the same for the two ows, this similarity conrms that
the k-! models predictions are consistent with the free-interaction
concept.
The numerical separation points for these ows are further
upstream than indicated by oil-ow measurements. Marshall and
Dolling [54] indicated that such ows include a low-frequency
separation-shock oscillation. Adams [55] found this oscillation in a
direct numerical simulation of a Mach 3 compression-corner ow.
The time-mean pressure distribution upstream of the corner is
affected by these oscillations, for which the frequency content
includes substantial energy at time scales of the mean motion. This
unsteadiness is responsible for the apparent mismatch between the
beginning of the pressure rise and the separation point. Because
computations with the k-! model are so close to measured properties,
yet display no low-frequency oscillation of the shock, we can
reasonably conclude that the computations effectively incorporate
the slow oscillation into the Favre-averaged ow variables.
Figure 18 indicates how the separation-point location for the
24 deg compression-corner ow varies with Clim . As shown, similar
to the effect for an incompressible backward-facing step (see
Fig. 13), xs increases monotonically as Clim increases. Selecting
Clim  7=8 yields a value of xs  1:82, which provides a very
close match to most details of this oweld.
Figure 19 shows that using Clim  1 produces a separation bubble
roughly twice the measured size. This explains why Menters [7]
model fares so poorly for this ow [56].

E.

Mach 11 Reecting Shock

We turn now to a hypersonic ow: namely, the Mach 11 shockwave/boundary-layer interaction investigated by Holden [57]. The
incident shock makes a 17.6 deg angle with the surface and increases
the static pressure by a factor of 70. The shock angle was adjusted
from the precise value used in the experiment to match the overall

Fig. 18 Effect of the stress-limiter coefcient on computed separationpoint location for Mach 3 ow into a 24 deg compression corner.

Fig. 19 Comparison of computed and measured surface pressure and skin friction for Mach 3 ow into a 24 deg compression corner: Menter [7]
k-!=k-" model (solid line), new k-! model with Clim  1 (dashed line), Settles et al. [50] (circles), and Dolling and Murphy [51] (squares).

2835

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

WILCOX

Fig. 20 Effect of the stress limiter on the new k-! model for a Mach 11
shock-wave/boundary-layer interaction: with limiter (solid line), without
limiter (dashed line), and Holden [57] (circles).

inviscid pressure rise for an assumed specic heat ratio   1:4. The
surface is highly cooled with a wall-to-adiabatic-wall temperature
ratio of Tw =Taw  0:2.
Figure 20 compares computed and measured surface pressure for
the new k-! model with and without the stress limiter. As shown, the
limiter increases separation-bubble length from 0.34 to 1:53o . The
computed surface-pressure rise is much closer to the measured rise
when the limiter is used. As with the Mach 3 applications of the
preceding subsection, this indicates that the predicted shock pattern
closely matches the experimental pattern. Holden [57] estimated the
size of the separation bubble to be about 1:00o . The surface-pressure
data suggest a separation bubble about twice that size.
As with all of the computations discussed in this paper, the
turbulent Prandtl number was chosen to be PrT  0:89. In general,
for this and other hypersonic shock-separated ows done with the
new k-! model (see Wilcox [14]), computed heat transfer at the
reattachment point is about 50% higher than measured. This
difculty is characteristic of turbulence models that base the
turbulent heat-ux vector on Reynolds analogy. As shown by Xiao
et al. [58], realistic predictions for hypersonic reattachment point
heat transfer can be achieved by constructing additional model
equations to compute the heat-ux vector.

V.

Conclusions

There are two signicant results of the research described in this

paper. First, only a small range of values for the cross-diffusion
coefcient d exists that yields satisfactory spreading rates for free
shear ows (see Figs. 1 and 2). Second, using too large of a value for
the stress-limiter strength Clim causes the k-! model to predict
separated regions much larger than measured for ows above
transonic speeds (see Fig. 19).
The new k-! model retains all of the strengths of previous models
developed by the author. Specically, the model is as accurate for
attached boundary layers, backward-facing steps, and mildly
separated incompressible ows. The original k-! model presented
by Saffman [2] had ve empirical closure coefcients. Of necessity,
some of the models elegance and simplicity was sacriced to
remove sensitivity to freestream boundary conditions and to achieve
more realistic predictions for free shear ows. The price was one
additional closure coefcient, d , and two empirical closure
functions [see Eqs. (1113)]. And, of course, the model requires a
replacement for the linear constitutive relation between Reynolds
stresses and mean strain rate used in the original model. The stress
limiter is the simplest such relationship available, and it adds just one
additional closure function [see Eq. (6)].
Inclusion of a cross-diffusion term in the ! equation 1) greatly
improves the models predictions for all ve of the classic free shear

ows (see Table 1) and 2) signicantly reduces the models

sensitivity to nite freestream boundary conditions on turbulence
parameters (see Fig. 4).
With inclusion of a stress limiter, the new k-! model predicts
reasonably close agreement with measured properties of shockseparated ows for transonic, supersonic, and hypersonic regimes.
Although discrepancies can be reduced even further by increasing
the strength of the limiter in specic cases (most notably for transonic
ows), choosing a limiter strength of Clim  7=8 appears to be the
optimum choice for covering the entire range of ow speeds from
incompressible to hypersonic.
The fact that all of the results presented in this paper were achieved
without any explicit compressibility modications to the k-! model
is entirely consistent with Morkovins [59] hypothesis. That is, the
effect of density uctuations on the turbulence is small provided they
remain small relative to the mean density. Although the model
predicts larger-than-measured values of heat ux at the reattachment
point in a hypersonic ow, that is an inaccuracy caused by a faulty
constitutive relation rather than a breakdown of Morkovins
hypothesis.
Although not discussed in this paper, the new k-! model fails to
match the measured reduction of spreading rate for a compressible
mixing layer. As discussed in great detail by Wilcox [14], density
uctuations for a compressible mixing layer are much larger than in
wall-bounded ows and are not small relative to the mean density.
Hence, Morkovins [59] hypothesis fails and the model, like all
turbulence models, will require ad hoc compressibility modications
to match measurements.

Appendix A: Numerical Accuracy

The computations presented in this paper were done with several
computer programs that apply to three different types of ows:
namely, free shear ows, attached boundary layers, and ows with
boundary-layer separation. All of these programs, including source
code, are included in the textbook by Wilcox [14]. The purpose of
this Appendix is to briey describe the programs and to document the
results of an iteration and grid-resolution study for the numerical
results discussed in this paper.
In all of the cases, the grid convergence index (GCI) devised by
Roache [60] is presented for the most sensitive ow property. This
index, based on generalized Richardson extrapolation, provides an
excellent measure of the computations accuracy. The ratio of negrid to coarse-grid dimensions is denoted by r. For example, a ne
grid with 1.5 times the number of points in the coarse grid would have
r  1:5. For a numerical method that is accurate to order p, the GCI
for a given ow property is
GCI  1:25

I.

j"h j
;
rp  1

"h 

fine  coarse
fine

(A1)

Free-Shear-Flow Programs

Three programs named JET, MIXER, and WAKE were used to

compute far-eld properties of jets, mixing layers, and wakes. All
three programs use time-marching methods to solve the nonlinear
two-point boundary-value problems attending use of the similaritysolution method for simple turbulent ows. The solution algorithm
used is based on implicit CrankNicolson [61] differencing. To
render straightforward and easy-to-modify programs, each equation

Table A1
Flow
Far wake
Mixing layer
Plane jet
Round jet
Radial jet

Grid-resolution results for spreading rate

Fine-grid size

0coarse

0fine

GCI

201
201
201
201
201

2
2
2
2
2

0.32640
0.09599
0.10780
0.09411
0.09914

0.32600
0.09643
0.10740
0.09388
0.09890

0.05%
0.19%
0.16%
0.10%
0.10%

2836

WILCOX

Table A2
Flow
Samuel and Joubert experiment [29]
Stratford [30]
Coles Mach 4.5 experiment [31]
Watson Mach 10.3 experiment [31]

Table A3

Grid-resolution results for skin friction

Figure

Fine-grid size

103
cf coarse

103
cf fine

GCI

6
6
7
7

201
201
201
201

2
2
2
2

1.30875
0.67697
1.21393
0.23706

1.31154
0.67203
1.21390
0.23741

0.09%
0.31%
0.00%
0.06%

Grid-resolution results for separated-region size

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

Flow
High Reynolds number incompressible backstep
Low Reynolds number incompressible backstep
Transonic bump
Mach 2 backstep
Mach 3 20 deg compression corner
Mach 3 24 deg compression corner
Mach 3 shock/boundary-layer interaction
Mach 11 shock/boundary-layer interaction

of a given turbulence model is solved independently using a standard

tridiagonal-matrix inversion algorithm.
An additional transformation devised by Rubel and Melnik [62]
was used in all three of the free-shear-ow programs that greatly
improves numerical accuracy. Because the transformation automatically stretches the transformed distance in regions of rapidly varying
ow properties, a grid with equally spaced points can be used.
Consequently, as validated by computations on three different nite
difference grids, the programs are exactly second-order-accurate
(i.e., p  2).
All computations were run until machine accuracy was achieved,
which assures that iteration convergence was attained. Table A1
summarizes the results of a grid-resolution study. Computed
spreading rate for the ve basic free shear ows is listed for nite
difference grids with 101 and 201 points. In general, the GCI is even
smaller for ow properties throughout the numerical oweld.

Figure

Fine-grid Size

xcoarse

xfine

GCI

11
12
14
15
17
17
17
19

301  163
201  163
201  101
401  201
401  201
401  201
401  201
500  300

1.25
1.25
1.25
1.25
1.41
1.41
1.41
1.19

7.158
7.451
0.414
2.728
1.151
2.785
4.603
1.519

7.070
7.280
0.410
2.672
1.160
2.825
4.650
1.530

3.1%
5.9%
2.6%
5.3%
1.1%
2.0%
1.5%
2.5%

maximum residual being reduced by 6 to 10 orders of magnitude.

Table A3 includes computed separation-bubble length
(x  xr  xs ) for the shock-separated ows and reattachment
length (x  xr ) for the backstep applications. As with the freeshear-ow and boundary-layer computations already discussed, the
GCI is even smaller for ow properties throughout the numerical
oweld. For example, the skin-friction and pressure coefcients
downstream of reattachment for the backward-facing steps have a
GCI of about 1%.

Acknowledgment
The research presented in this paper was totally funded by DCW
Industries, Inc.

References
II.

Attached Boundary Layers

A program named EDDYBL was used for attached boundary

layers. The program applies to attached, compressible, twodimensional, and axisymmetric boundary layers. It includes the new
k-! model as well as many popular algebraic, one-equation, and twoequation models.
Program EDDYBL uses the Blottner [63] variable-grid method
augmented with an algorithm devised by Wilcox [64] to permit large
streamwise steps. The program uses adaptive gridding in the
streamwise direction, decreasing step size as the number of iterations
needed for convergence increases and vice versa. Computations on
three different grids show that the effective order of accuracy of the
numerical algorithm is p  1:9. Wilcox [14] provided an in-depth
discussion of the algorithm.
Table A2 summarizes results of a grid-resolution study. The table
includes computed skin friction at the last streamwise station for the
four attached boundary-layer cases shown in Figs. 6 and 7. As with
the free-shear-ow computations already discussed, the GCI is even
smaller for other ow properties throughout the numerical oweld.
III.

Flows with Boundary-Layer Separation

A program named EDDY2C was used for ows with boundarylayer separation. The program uses the MacCormack [41] fully
implicit ux-splitting method with GaussSeidel line relaxation.
Computations on numerous ows with three different grids show
that the effective order of accuracy of the numerical algorithm is
typically p  1:8.
The ow property that generally takes longest to converge is the
size of the separated region. All computations reported in this paper
were run long enough to insure iteration convergence, with the

[1] Kolmogorov, A. N., Equations of Turbulent Motion of an

Incompressible Fluid, Izvestiya Academii Nauk USSR: Physics,
Vol. 6, Nos. 12, 1942, pp. 5658.
[2] Saffman, P. G., A Model for Inhomogeneous Turbulent Flow,
Proceedings of the Royal Society of London, Series A: Mathematical
and Physical Sciences, Vol. 317, No. 1530, 1970, pp. 417433.
doi:10.1098/rspa.1970.0125
[3] Wilcox, D. C., and Alber, I. E., A Turbulence Model for High Speed
Flows, Proceedings of the 1972 Heat Transfer and Fluid Mechanics
Institute, Stanford Univ. Press, Stanford, CA, 1972, pp. 231252.
[4] Wilcox, D. C., Reassessment of the Scale Determining Equation for
Advanced Turbulence Models, AIAA Journal, Vol. 26, No. 11, 1988,
pp. 12991310.
doi:10.2514/3.10041
[5] Wilcox, D. C., Turbulence Modeling for CFD, 2nd ed., DCW
Industries, Inc., La Caada, CA, 1998.
[6] Speziale, C. G., Abid, R., and Anderson, E. C., A Critical Evaluation of
Two-Equation Models for Near Wall Turbulence, AIAA Paper 901481, Seattle, WA, 1990.
[7] Menter, F. R., Improved Two-Equation k-! Turbulence Models for
Aerodynamic Flows, NASA TM-103975, 1992.
[8] Wilcox, D. C., A Two-Equation Turbulence Model for Wall-Bounded
and Free-Shear Flows, AIAA Paper 93-2905, Orlando, FL, 1993.
[9] Kok, J. C., Resolving the Dependence on Freestream Values for the
k-! Turbulence Model, AIAA Journal, Vol. 38, No. 7, 2000, pp. 1292
1295.
[10] Hellsten, A., New Advanced k-! Turbulence Model for High-Lift
Aerodynamics, AIAA Journal, Vol. 43, No. 9, 2005, pp. 18571869.
doi:10.2514/1.13754
[11] Coakley, T. J., Turbulence Modeling Methods for the Compressible
Navier-Stokes Equations, AIAA Paper 83-1693, 1983.
[12] Huang, P. G., Physics and Computations of Flows with Adverse
Pressure Gradients, Modeling Complex Turbulent Flows, Kluwer,
Dordrecht, The Netherlands, 1999, pp. 245258.

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

WILCOX

[13] Kandula, M., and Wilcox, D. C., An Examination of k-! Turbulence

Model for Boundary Layers, Free Shear Layers and Separated Flows,
AIAA Paper 95-2317, San Diego, CA, 1995.
[14] Wilcox, D. C., Turbulence Modeling for CFD, 3rd ed., DCW Industries,
Inc., La Caada, CA, 2006.
[15] Papp, J. L., and Dash, S. M., Turbulence Model Unication and
Assessment for High-Speed Aeropropulsive Flows, AIAA
Paper 2001-0880, 2001.
[16] Schlichting, H., and Gersten, K., Boundary Layer Theory, 8th ed.,
SpringerVerlag, Berlin, 1999.
[17] Andersen, P. S., Kays, W. M., and Moffat, R. J., The Turbulent
Boundary Layer on a Porous Plate: An Experimental Study of the Fluid
Mechanics for Adverse Free-Stream Pressure Gradients, Dept. of
Mechanical Engineering, Stanford Univ., Rept. HMT-15, Stanford,
CA, 1972.
[18] Peng, S.-H., Davidson, L., and Holmberg, S., A Modied LowReynolds Number k-! Model for Recirculating Flows, Journal of
Fluids Engineering, Vol. 119, No. 4, 1997, pp. 867875.
doi:10.1115/1.2819510
[19] Launder, B. E., and Sharma, B. I., Application of the Energy
Dissipation Model of Turbulence to the Calculation of Flow Near a
Spinning Disc, Letters in Heat and Mass Transfer, Vol. 1, No. 2, 1974,
pp. 131138.
doi:10.1016/0094-4548(74)90150-7
[20] Lele, S. K., A Consistency Condition for Reynolds Stress Closures,
Physics of Fluids Vol. 28, 1985, p. 64.
doi:10.1063/1.865126
[21] Pope, S. B., An Explanation of the Turbulent Round-Jet/Plane-Jet
Anomaly, AIAA Journal, Vol. 16, No. 3, 1978, pp. 279281.
doi:10.2514/3.7521
[22] Rubel, A., On the Vortex Stretching Modication of the k-"
Turbulence Model: Radial Jets, AIAA Journal, Vol. 23, No. 7, 1985,
pp. 11291130.
doi:10.2514/3.9051
[23] Tanaka, T., and Tanaka, E., Experimental Study of a Radial Turbulent
Jet, Bulletin of the JSME, Vol. 19, No. 133, 1976, pp. 792799.
[24] Witze, P. O., and Dwyer, H. A., The Turbulent Radial Jet, Journal of
Fluid Mechanics, Vol. 75, 1976, pp. 401417.
doi:10.1017/S0022112076000293
[25] Yakhot, V., and Orszag, S. A., Renormalization Group Analysis of
Turbulence, 1: Basic Theory, Journal of Scientic Computing, Vol. 1,
No. 1, 1986, pp. 351.
doi:10.1007/BF01061452
[26] Fage, A., and Falkner, V. M., Note on Experiments on the Temperature
and Velocity in the Wake of a Heated Cylindrical Obstacle,
Proceedings of the Royal Society of London, Series A: Mathematical
and Physical Sciences, Vol. 135, 1932, pp. 702705.
[27] Weygandt, J. H., and Mehta, R. D., Three-Dimensional Structure of
Straight and Curved Plane Wakes, Journal of Fluid Mechanics,
Vol. 282, 1995, p. 279.
doi:10.1017/S0022112095000140
[28] Bardina, J., Huang, P. G., and Coakley, T. J., Turbulence Modeling
Validation, AIAA Paper 97-2121, 1997.
[29] Kline, S. J., Cantwell, B. J., and Lilley, G. M. (eds.), Complex Turbulent
Flows, Stanford Univ., Stanford, CA, 1981.
[30] Stratford, B. S., An Experimental Flow with Zero Skin Friction
Throughout its Region of Pressure Rise, Journal of Fluid Mechanics,
Vol. 5, 1959, pp. 1735.
doi:10.1017/S0022112059000027
[31] Fernholz, H. H., and Finley, P. J., A Further Compilation of
Compressible Boundary Layer Data with a Survey of Turbulence
Data, AGARD AGARDograph 263, Neuilly-sur-Seine, France, 1981.
[32] Klebanoff, P. S., Characteristics of Turbulence in a Boundary Layer
with Zero Pressure Gradient, NACA TN 1247, 1955.
[33] Wieghardt, K., and Tillman, W., On the Turbulent Friction Layer for
Rising Pressure, NACA TM 1314, 1951.
[34] Winter, K., and Gaudet, L., Turbulent Boundary-Layer Studies at High
Reynolds Number at Mach Numbers Between 0.2 and 2.8,
Aeronautical Research Council, Research and Memorandum
No. 3712, London, 1973.
[35] Durbin, P. A., On the k-" Stagnation Point Anomaly, International
Journal of Heat and Fluid Flow, Vol. 17, No. 1, 1996, pp. 8990.
doi:10.1016/0142-727X(95)00073-Y
[36] Moore, J. G., and Moore, J., Realizability in Two-Equation
Turbulence Models, AIAA Paper 99-3779, 1999.
[37] Bradshaw, P., Ferriss, D. H., and Atwell, N. P., Calculation of
Boundary Layer Development Using the Turbulent Energy Equation,
Journal of Fluid Mechanics, Vol. 28, Pt. 3, 1967, pp. 593616.
doi:10.1017/S0022112067002319

2837

[38] Saffman, P. G., and Wilcox, D. C., Turbulence-model Predictions for

Turbulent Boundary Layers, AIAA Journal, Vol. 12, No. 4, 1974,
pp. 541546.
doi:10.2514/3.49282
[39] Wilcox, D. C., and Traci, R. M., Analytical Study of Freestream
Turbulence Effects on Stagnation Point Flow and Heat Transfer,
AIAA Paper 74-515, 1974.
[40] Bearman, P. W., Some Measurements of the Distortion of Turbulence
Approaching a Two-Dimensional Bluff Body, Journal of Fluid
Mechanics, Vol. 53, Part 3, 1972, pp. 451467.
doi:10.1017/S0022112072000254
[41] MacCormack, R. W., Current Status of Numerical Solutions of the
Navier-Stokes Equations, AIAA Paper 85-32, 1985.
[42] Driver, D. M., and Seegmiller, H. L., Features of a Reattaching
Turbulent Shear Layer in Divergent Channel Flow, AIAA Journal,
Vol. 23, No. 2, 1985, pp. 163171.
doi:10.2514/3.8890
[43] Jovic, S., and Driver, D., Reynolds Number Effect on the Skin Friction
in Separated Flows Behind a Backward-Facing Step, Experiments in
Fluids, Vol. 18, No. 6, 1995, pp. 464467.
doi:10.1007/BF00208471
[44] Jovic, S., and Driver, D., Backward-Facing Step Measurements at
Low Reynolds Number, NASA TM-108870, 1994.
[45] Bachalo, W. D., and Johnson, D. A., An Investigation of Transonic
Turbulent Boundary Layer Separation Generated on an Axisymmetric
Flow Model, AIAA Paper 79-1479, 1979.
[46] Spalart, P. R., and Allmaras, S. R., A One-Equation Turbulence Model
for Aerodynamic Flows, AIAA Paper 92-439, 1992; also La
Recherche Aerospatiale : bulletin Bimestriel de lOfce National
dEtudes et de Recherches Aerospatiales, No. 1, 1994, p. 5.
[47] Samimy, M., Petrie, H. L., and Addy, A. L., Reattachment and
Redevelopment of Compressible Turbulent Free Shear Layers,
International Symposium on Laser Anemometry, Vol. 33, American
Society of Mechanical Engineers, Fluids Engineering Div., New York,
1985, pp. 159166.
[48] Petrov, G., Likhusin, V., Nekrasov, I., and Sorkin, L., Inuence of
Viscosity on the Supersonic Flow with Shock Waves, Central Institute
of Aviation Motors, No. 224, 1952 [in Russian].
[49] Chapman, D., Kuehn, D., and Larson, H., Investigation of Separated
Flows in Supersonic and Subsonic Streams with Emphasis on the Effect
of Transition, NACA Rept. 1356, 1957.
[50] Settles, G. S., Vas, I. E., and Bogdonoff, S. M., Details of a Shock
Separated Turbulent Boundary Layer at a Compression Corner, AIAA
Journal, Vol. 14, No. 12, 1976, pp. 17091715.
doi:10.2514/3.61513
[51] Dolling, D. S., and Murphy, M. T., Unsteadiness of the Separation
Shock-wave Structure in a Supersonic Compression Ramp Floweld,
AIAA Journal, Vol. 21, No. 12, 1983, pp. 16281634.
doi:10.2514/3.60163
[52] Reda, D. C., and Murphy, J. D., Shock-wave/Turbulent-BoundaryLayer Interactions in Rectangular Channels, AIAA Journal, Vol. 11,
No. 2, 1973, pp. 139140.
doi:10.2514/3.50445also AIAA Paper 72-715, 1972.
[53] Murthy, V. S., and Rose, W. C., Wall Shear Stress Measurements in a
Shock-Wave/Boundary-Layer Interaction, AIAA Journal, Vol. 16,
No. 7, 1978, pp. 667672.
doi:10.2514/3.60956
[54] Marshall, T. A., and Dolling, D. S., Computation of Turbulent,
Separated, Unswept Compression Ramp Interactions, AIAA Journal,
Vol. 30, No. 8, 1992, pp. 20562065.
doi:10.2514/3.11179
[55] Adams, N., Direct Simulation of the Turbulent Boundary Layer along
a Compression Ramp at M  3 and Re  1685, Journal of Fluid
Mechanics, Vol. 420, 2000, pp. 4783.
doi:10.1017/S0022112000001257
[56] Forsythe, J. R., Strang, W. Z., and Hoffmann, K. A., Validation of
Several Reynolds-Averaged Turbulence Models in a 3-D Unstructured
Grid Code, AIAA Paper 00-2552, 2000.
[57] Holden, M. S., A Study of Flow Separation in Regions of Shock WaveBoundary Layer Interaction in Hypersonic Flow, AIAA Paper 781168, 1978.
[58] Xiao, X., Edwards, J. R., Hassan, H. A., and Gaffney, R. L., Jr., Role of
Turbulent Prandtl Number on Heat Flux at Hypersonic Mach
Numbers, AIAA Paper 2005-1098, 2005.
[59] Morkovin, M. V., Effects of Compressibility on Turbulent Flow, The
Mechanics of Turbulence, edited by A. Favre, Gordon and Breach, New
York, 1962, p. 367.
[60] Roache, P. J., Verication and Validation in Computational Science
and Engineering, Hermosa, Albuquerque, NM, 1998.

2838

WILCOX

Downloaded by University of Newcastle on March 12, 2014 | http://arc.aiaa.org | DOI: 10.2514/1.36541

[61] Crank, J., and Nicolson, P., A Practical Method for Numerical
Evaluation of Solutions of Partial Differential Equations of the HeatConduction Type, Proceedings of the Cambridge Philosophical
Society (Mathematical and Physical Sciences), Vol. 43, No. 50, 1947,
pp. 5067.
[62] Rubel, A., and Melnik, R. E., Jet, Wake and Wall Jet Solutions Using a
k-" Turbulence Model, AIAA Paper 84-1523, 1984.
[63] Blottner, F. G., Variable Grid Scheme Applied to Turbulent Boundary
Layers, Computer Methods in Applied Mechanics and Engineering,

Vol. 4, No. 2, 1974, pp. 179194.

doi:10.1016/0045-7825(74)90033-4
[64] Wilcox, D. C., Algorithm for Rapid Integration of Turbulence Model
Equations on Parabolic Regions, AIAA Journal, Vol. 19, No. 2, 1981,
pp. 248251.
doi:10.2514/3.7766

P. Givi
Associate Editor